Development and application of the Kalman filter ... - IRTG Heidelberg
Development and application of the Kalman filter ... - IRTG Heidelberg
Development and application of the Kalman filter ... - IRTG Heidelberg
You also want an ePaper? Increase the reach of your titles
YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.
<strong>Development</strong> <strong>and</strong> <strong>application</strong><br />
<strong>of</strong> <strong>the</strong> <strong>Kalman</strong> <strong>filter</strong> method<br />
in <strong>the</strong> CBM <strong>and</strong> ALICE experiments<br />
Sergey Gorbunov<br />
GSI / KIP<br />
<strong>Heidelberg</strong><br />
October 19, 2007
Outline<br />
• Fit problem in event reconstruction in HEP<br />
• The <strong>Kalman</strong> Filter method<br />
• <strong>Kalman</strong> Filter for on-line event reconstruction in CBM<br />
• KF track fit in CBM<br />
• CA+KF tracker in CBM<br />
• Fast SIMDized <strong>Kalman</strong> Filter<br />
• <strong>Kalman</strong> Filter for on-line event reconstruction in ALICE HLT<br />
• CA tracker for ALICE HLT<br />
• KF fit with 3D track model<br />
• Reconstruction <strong>of</strong> vertices <strong>and</strong> decayed particles<br />
• Summary<br />
<strong>Heidelberg</strong>, 19.10.2007 Sergey Gorbunov, GSI/KIP 2/15
Fit problem in event reconstruction<br />
• Track finding<br />
• Track fitting<br />
• Track merging<br />
• Reconstruction <strong>of</strong> vertices<br />
<strong>and</strong> decayed particles<br />
D 0 K -<br />
<strong>Heidelberg</strong>, 19.10.2007 Sergey Gorbunov, GSI/KIP 3/15<br />
ππ ++
The <strong>Kalman</strong> <strong>filter</strong> method<br />
r 0<br />
initial<br />
track<br />
m 1<br />
σ 1<br />
σ m1<br />
r 1<br />
m 2<br />
<strong>Heidelberg</strong>, 19.10.2007 Sergey Gorbunov, GSI/KIP 4/15<br />
r 2<br />
m 3<br />
r 3<br />
fitted track<br />
The <strong>Kalman</strong> <strong>filter</strong> is <strong>the</strong> most used method for <strong>the</strong> fit problem.<br />
The method is well known in HEP, but in modern experiments<br />
<strong>the</strong> large amount <strong>of</strong> data <strong>and</strong> specific problems require development <strong>of</strong> new<br />
<strong>Kalman</strong> <strong>filter</strong>-based algorithms as well as investigations <strong>of</strong> <strong>the</strong> basic method.
KF fit in <strong>the</strong> CBM experiment<br />
The full set <strong>of</strong> fitting utilities<br />
has been developed<br />
Tracking in STS<br />
Tracking in muon chambers<br />
Tracking in TRD<br />
Track-ring match in RICH<br />
Hit ga<strong>the</strong>ring (MVD)<br />
Global tracking<br />
Reconstruction <strong>of</strong> vertices<br />
<strong>and</strong> decayed particles<br />
Smoo<strong>the</strong>r utility<br />
CBM reconstruction challenge<br />
~ 1000 charged particles/collision<br />
85% fake hits<br />
non-homogeneous magnetic field<br />
10 7 Au+Au collisions/sec<br />
no simple trigger<br />
CbmKF is <strong>the</strong> default fitting package in CBM<br />
<strong>Heidelberg</strong>, 19.10.2007 Sergey Gorbunov, GSI/KIP 5/15
CA+KF tracker for CBM<br />
CA w/o KF CA with KF<br />
Efficiency, %<br />
95.29<br />
90.52<br />
76.05<br />
0.00<br />
2.18<br />
Track category<br />
Reference set<br />
All set<br />
Extra set<br />
Clone<br />
Ghost<br />
Efficiency, %<br />
99.45<br />
96.98<br />
89.46<br />
<strong>Heidelberg</strong>, 19.10.2007 Sergey Gorbunov, GSI/KIP 6/15<br />
0.01<br />
0.61
Speed up <strong>of</strong> <strong>the</strong> <strong>Kalman</strong> <strong>filter</strong><br />
Problem: Pixel geometry (1 sec ) => double-sided strips (3 min)<br />
Reco time = N combinations*fit time => speed up <strong>of</strong> KF fit utilities needed<br />
Stage<br />
0<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
Initial scalar version<br />
Approximation <strong>of</strong> <strong>the</strong> magnetic field<br />
Optimization <strong>of</strong> <strong>the</strong> algorithm<br />
SIMDization<br />
Porting to SPE<br />
Parallelization on 16 SPE’s<br />
Final SIMDized version on Cell<br />
FPGA<br />
Description<br />
Time/track<br />
12 ms<br />
240 µs<br />
7.2 µs<br />
1.6 µs<br />
1.1 µs<br />
0.1 µs<br />
0.1 µs<br />
? ns<br />
Speedup<br />
120000<br />
<strong>Heidelberg</strong>, 19.10.2007 Sergey Gorbunov, GSI/KIP 7/15<br />
-<br />
50<br />
35<br />
4.5<br />
1.5<br />
10<br />
?<br />
Dual Cell Blade<br />
at <strong>the</strong> IBM Laboratory,<br />
Böblingen, Germany<br />
Future,<br />
but realistic
KF fit on different architectures<br />
Fit <strong>of</strong> a single track:<br />
lxg1411<br />
eh102<br />
blade11bc4<br />
S.Gorbunov, U.Kebschull, I.Kisel, V.Lindenstruth, W.F.J.Müller, Fast SIMDized <strong>Kalman</strong> <strong>filter</strong> based track fit, acc. by CPC<br />
Fit <strong>of</strong> thous<strong>and</strong>s <strong>of</strong> tracks:<br />
<strong>Heidelberg</strong>, 19.10.2007 Sergey Gorbunov, GSI/KIP 8/15
Speed up <strong>of</strong> CA+KF tracker<br />
1000 times<br />
faster<br />
Central events Minimum bias events<br />
Efficiency, %<br />
97.04<br />
93.26<br />
79.95<br />
1.15<br />
2.28<br />
523<br />
78 ms<br />
All set<br />
Clone<br />
Ghost<br />
Track category<br />
Reference set (>1 GeV/c)<br />
Extra set (
CA+KF tracker for ALICE HLT<br />
All Set<br />
(Hits >10, P > .05)<br />
Eff Clone<br />
97.4<br />
14.2<br />
Reference Set<br />
(All Set + P>1.)<br />
Eff Clone<br />
99.5<br />
0.00<br />
Extra Set<br />
(All Set + P
CA+KF algorithm (sector view)<br />
step 0: Clusters<br />
YX view<br />
ZX view<br />
step 1: Cells<br />
YX view<br />
step 2: Neighbors step 3: Tracks ( KF )<br />
• No vertex constraint => physics, cosmics, beam gas, laser tracks…<br />
• Simple code, small memory => Cell, GPU,…<br />
• 3D track model r [8] = ( x, y, z, p x , p y , p z , q/P )<br />
ZX view<br />
ZX view<br />
<strong>Heidelberg</strong>, 19.10.2007 Sergey Gorbunov, GSI/KIP 11/15
Reconstruction <strong>of</strong> vertices <strong>and</strong> decayed particles<br />
D 0<br />
K -<br />
ππ ++<br />
KFParticle: particle parameters<br />
r [8] = ( x y z, p x p y p z , E, S(=L/p) )<br />
with full covariance matrix C [ 8×8 ]<br />
• Physical track model for mo<strong>the</strong>r [D 0 ] <strong>and</strong> daughters [ππ ++ , K - ]<br />
• State vector contains all <strong>the</strong> particle parameters<br />
both at decay <strong>and</strong> production vertices<br />
• Algorithm is suitable for vertex fit ( including primary vertex fit )<br />
• User friendly interface: class CbmKFParticle / AliKFParticle ;<br />
<strong>Heidelberg</strong>, 19.10.2007 Sergey Gorbunov, GSI/KIP 12/15
CbmKFParticle performance for D 0 fit in CBM<br />
D 0<br />
10 6 events<br />
Accuracy<br />
Pull<br />
Production Vertex [µm]<br />
x<br />
0.81<br />
1.14<br />
y<br />
0.73<br />
1.10<br />
z<br />
5.50<br />
1.11<br />
x<br />
2.64<br />
1.13<br />
K -<br />
ππ ++<br />
Decay Vertex [µm]<br />
y<br />
2.64<br />
1.13<br />
• Fixed target geometry<br />
• Non-uniform magnetic field<br />
• ~500 primary tracks<br />
63.88<br />
<strong>Heidelberg</strong>, 19.10.2007 Sergey Gorbunov, GSI/KIP 13/15<br />
z<br />
1.10<br />
P [%]<br />
0.79<br />
1.20<br />
Physical Parameters<br />
M [MeV/c]<br />
11.34<br />
1.19<br />
L [µm]<br />
64.10<br />
1.11<br />
cT [µm]<br />
9.81<br />
1.11
AliKFParticle performance for D 0 fit in ALICE<br />
10 6 events<br />
Accuracy<br />
Pull<br />
x<br />
61.89 49.07<br />
0.96 0.95<br />
D 0<br />
Production Vertex [µm]<br />
y<br />
60.95 48.79<br />
0.98 0.95<br />
z<br />
87.52 67.42<br />
1.01 0.98<br />
x<br />
100.6 75.41<br />
0.97 0.92<br />
Decay Vertex [µm]<br />
y<br />
K -<br />
75.03 97.4<br />
0.97 0.92<br />
116.1 88.6<br />
• Collider geometry<br />
• Uniform magnetic field<br />
• ~5 primary tracks<br />
<strong>Heidelberg</strong>, 19.10.2007 Sergey Gorbunov, GSI/KIP 14/15<br />
ππ ++<br />
z<br />
1.00 0.97<br />
V 0 fit time: 50 µs => HLT<br />
P [%]<br />
0.78 0.75<br />
0.94 0.92<br />
Physical Parameters<br />
M [MeV/c]<br />
10.4 9.9<br />
0.95 0.94<br />
CbmKFParticle = AliKFParticle => same code!<br />
L [µm]<br />
187.3 165.5<br />
0.98 0.93<br />
cT [µm]<br />
115.3 100.4<br />
0.98 0.93
Summary<br />
• Developed <strong>and</strong> implemented:<br />
The <strong>Kalman</strong> <strong>filter</strong> utilities for CBM reconstruction [ σP = 1.2% ]<br />
Fast SIMDized <strong>Kalman</strong> <strong>filter</strong> for CBM on-line reconstruction [ 1.6 µs / 0.1 µs ]<br />
CA+KF tracker for CBM [ Eff = 97%, 78 ms / 5 ms ]<br />
CA+KF tracker for ALICE HLT [ Eff = 99%, 100 ms ]<br />
Reconstruction <strong>of</strong> decayed particles for CBM <strong>and</strong> ALICE [σcT= 9.8 µm/100 µm, 50 µs]<br />
• Related publications :<br />
• S.Gorbunov, I.Kisel, Analytic formula for track extrapolation in non-homogeneous magnetic field, NIM A 559, (2006)<br />
• S.Gorbunov, U.Kebschull, I.Kisel, V.Lindenstruth, W.F.J.Müller, Fast SIMDized <strong>Kalman</strong> <strong>filter</strong> based track fit, CPC, (2007)<br />
• S.Gorbunov, I.Kisel, Primary vertex fit based on <strong>the</strong> <strong>Kalman</strong> <strong>filter</strong>, CBM-SOFT-note-2006-001, (2006)<br />
• S.Gorbunov, I.Kisel, Secondary vertex fit based on <strong>the</strong> <strong>Kalman</strong> <strong>filter</strong>, CBM-SOFT-note-2006-002, (2006)<br />
• S.Gorbunov, I.Kisel, Reconstruction <strong>of</strong> decayed particles based on <strong>the</strong> <strong>Kalman</strong> <strong>filter</strong>, CBM-SOFT-note-2007-003, (2007)<br />
• S.Gorbunov, I.Kisel, I.Vassiliev, Analysis <strong>of</strong> D0 detection in Au+Au collisions at 25 AGeV, CBM-PHYS-note-2005-001, (2005)<br />
• S.Gorbunov, I.Kisel, Elastic net for st<strong>and</strong>-alone RICH ring finding, NIM A 559, (2006)<br />
• Future plans (ALICE HLT):<br />
Track fit investigation<br />
Speed up <strong>and</strong> SIMDization<br />
Run <strong>the</strong> reconstruction core on <strong>the</strong> Cell processor <strong>and</strong> GPU<br />
Test on heavy ion events<br />
<strong>Heidelberg</strong>, 19.10.2007 Sergey Gorbunov, GSI/KIP 15/15